Modelling of Vehicle Longitudinal Dynamics using System

Identification
H. Syahira1, M. Abdullah1*, S. Ahmad1, M.A.S. Zainuddin1,2, and K.A. Tofrowaih3
1
Dept. of Mechanical and Aerospace Engineering, International Islamic University Malaysia, Jalan Gombak,
53100, Kuala Lumpur, Malaysia
2
Dept. of Automotive Engineering Technology, Kolej Kemahiran Tinggi MARA, Masjid Tanah, 78300, Melaka,
Malaysia
3
Dept. of Mechanical and Manufacturing Engineering, University Teknikal Malaysia Melaka, Hang Tuah Jaya,
76100, Melaka, Malaysia
*
[email protected]

Abstract: This paper presents an alternative modelling technique known as system identification to represent longitudinal
vehicle dynamics that relates input of pedals pressing to the output speed. The input and output data are recorded by
simulating a high-fidelity nonlinear vehicle mathematical model with suitable sampling time. Different excitation input
ranges are tested for the training process to get the best model representation. Once a suitable transfer function is obtained,
a PID controller is tuned and implemented to control the brake and throttle pedals of the car based on the desired velocity
profile. The simulation results show that the tuned PID controller managed to track the standard FTP75 drive cycles with an
acceptable response. Based on these findings, it is noted that the system identification model can become a good alternative
modelling technique to design a model-based controller or tune a classical controller for future application.

Nevertheless, to use the system identification modelling

1. Introduction approach, good data measurement is needed to capture the
Adaptive Cruise Control (ACC) is a system that can core dynamics of a system. It is noted that the model’s
automatically track and maintain a vehicle’s speed without accuracy is affected by the consistency of the measured data,
any human intervention. In the presence of lead vehicle, this and thus a good experimental design is needed. Dias [9]
system can automatically maintain a safe following distance shows that with a proper experimental design, the identified
by adjusting the speed via braking with the help of distance model for a car provides a reliable representation compared
sensor (radar or LIDAR) and a proper control algorithm. This to the actual behavior. Besides that, many successful system
system has become a standard feature in most modern cars by identification applications have been reported since the 1960s
offering several attractive benefits, such as reducing leg ranging from electrical, mechanical, and thermal applications
fatigue, fuel consumption, driving errors and road accidents [10]. From a control perspective, system identification is
[1-3]. According to the Society of Automotive Engineering much more practical as one does not need to derive a set of
(SAE), the ACC system is considered one of the basic differential equations rather than measuring data from the
features that is required for level 2 autonomous driving experimental design and choosing the proper model structure
classification. to reach the model’s accuracy.
One of the challenges in designing the ACC system is to Nevertheless, there are many choices of the model that can
develop a responsive and optimal control action to track the be used, and a user also needs to select a proper order for the
desired speed [3,4]. If the response is too fast, it may lead to model. Using a higher-order model with more parameters is
an uncomfortable driving experience, while if it is too slow, not always a better option as it can result in the modelling of
it may worsen the traffic flow [5]. Hence, it is necessary to nonexistence dynamics and noise characteristics [11].
tune the controller correctly, where a reliable mathematical Conversely, if the order of a model is too low, it will fail to
model is needed to represent the longitudinal vehicle capture the core dynamics of a system. Besides, system
dynamics [6]. However, deriving the equation of motion of a identification may be simple when given a linear system, but
vehicle requires a high level of understanding regarding the it can be difficult in practice due to noisy data and
underlying dynamics and accurate modelling parameters [7]. nonlinearities [11]. Thus, before the modelling process, a
For a complex system like engine or power train, a lot of proper testing methodology concerning signal processing is
simplification and assumption are applied to reduce the needed.
complexity of the model. Since a simplified mathematical When dealing with vehicle dynamics, by nature, the
model is enough to design and tune any general controller [6], nonlinearity is very high due to the changing in gear ratio and
an alternative modelling technique known as system driving speed [7]. Besides, there is a physical constraint in
identification can be utilized to represent the system both actuators where it cannot exceed 100% of pedal
dynamics. This method constructs a model based on pressing. This work will present a proper analysis of the
measurement inputs and outputs from an actual process, system identification approach, along with some
which may be more convenient than the previous method [8]. investigation of the impact of different data ranges on the

Authorized licensed use limited to: Purdue University. Downloaded on February 18,2025 at 15:41:31 UTC from IEEE Xplore. Restrictions apply.
1
model accuracy. Once a suitable model is developed, a
control algorithm can be designed or tuned based on the
identified model for the ACC application. This paper’s
structure is organized as follows: Section 2 will discuss the
methodology for system identification, Section 3 presents the
simulation results, and Section 4 provides the conclusion.

2. Methodology
2.1. Mathematical Model for the Plant

Since this work does not involve experimentation with an

actual vehicle, a high-fidelity mathematical model is used as Fig. 3. Wheels and brake block.
a plant, as shown in Fig. 1. The block diagram consists of Several types of braking systems can be used for this block.
three major blocks: the driver, wheels with brakes, and In this work, a disc brake system is selected where it will
vehicle dynamics. Each block has its mathematical equations convert the brake cylinder pressure into a force. The force
that describe the respective dynamics. This model is obtained will be applied to the brake pad and clamped to the disc to
from the MATLAB vehicle dynamics course, which is slow down or stop the vehicle completely. To calculate the
available in [12]. brake torque, a piecewise linear equation is used, where:

μ௉గ஻ೌమ ோ೘ ே೛ೌ೏ೞ
‫ ് ܰ ݄݊݁ݓ‬0
ସ
ܶ=ቐ (2)
μೞ೟ೌ೟೔೎ ௉గ஻ೌమ ே೛ೌ೏ೞ
‫ = ܰ ݄݊݁ݓ‬0
ସ

The parameters description of Eq. (2) is given in Table 1. In

Fig. 1. Block diagram of the plant.
general, when the car is not moving, the brake torque will use
Inside the driver block, there are two inputs to the overall the coefficient of static friction, and as the vehicle moves, the
system: the percentage of throttle and brake pedal positions. coefficient of kinetic friction will be used. Thus, if the car is
It should be noted that the power train dynamic for this plant not moving, no matter how large the force is applied, the
is simplified to a first-order transfer function that relates the vehicle’s velocity will be zero.
pedal position to torque command, as shown in Fig. 2 A magic formula constant values are used inside the tire
(APPDynamics). In reality, this relationship will be more Table 1 Description of the parameters in Eq. (2).
complex as many parameters need to be considered, such as Parameters Description
changing the gear ratio, differential slip, coefficient of
friction and others. Nevertheless, since the main scope of this T Brake torque
project is focusing on integrating the braking system with the P Applied brake pressure
speed control, the available Simulink block is adequate for N Wheel speed
this task. Npads Number of brake pads
μstatic Disk pad coefficient of static friction
μ Disk pad coefficient of kinetic friction
Ba Brake actuator bore diameter
Rm Mean radius of brake pad force
application on the brake rotor
Ra The outer radius of the brake pad
Ri The inner radius of the brake pad
Fig. 2. Driver block diagram.
block to model the longitudinal wheel as given in Eq. (3). The
The second block is the wheels and brakes. This block block implements the magic formula as a specific form of the
consists of wheels and a braking system, as shown in Fig. 3. tire characteristic function, characterized by four
This block calculates the inertial response of the wheel dimensionless coefficients (B, C, D, E) representing the
subject to axle losses, brake and drives torque, tire rolling stiffness, shape, peak, and curvature. The coefficients are
resistance, and ground contact, as given in Eq. (1). The input selected based on default empirical tire data for common road
torque, Ti is the summation of the applied axle torque, Ta, conditions as specified in the toolbox.
braking torque Tb, and moment from the combined tire
torque, Td. ‫ି݊ܽݐ ܥ(݊݅ݏܦݖܨ = ݔܨ‬ଵ [{‫ ݇ܤ‬െ ‫ି݊ܽݐ ݇ܤ[ܧ‬ଵ (‫]}])݇ܤ‬ (3)

ܶ݅ = ܶܽ െ ܾܶ + ܶ݀ (1) The third block is vehicle dynamics, as shown in Fig. 4,

which consists of the summation between traction, resistance,
and aerodynamics forces. Newton’s second law is used to
derive the equation of motion as given in Eq. (4). The
2

Authorized licensed use limited to: Purdue University. Downloaded on February 18,2025 at 15:41:31 UTC from IEEE Xplore. Restrictions apply.
 మ
parameters Ft is a traction force, Fr is the resistance force, ቚห௫ି௫ೝ೐೑ หቚ
and Fd is an aerodynamic force. Details explanation %݂݅‫= ݐ‬ (5)
ேೞ
regarding this block can be found in the MATLAB
documentation [12]. 2.3. PID Controller

‫ ݐܨ‬െ ݉݃ ‫ ݊݅ݏ‬Ʌ െ ‫ ݎܨ‬െ ‫ܽ݉ = ݀ܨ‬ (4) Once a simplified transfer function is obtained, a PID
controller will be tuned to control the pedal position of both
braking and throttle based on the desired speed. Fig. 6 shows
a schematic diagram where PID is implemented with the
identified transfer function. In this process, the three PID
gains, which are proportional Kp, integral Ki, and derivative
Kd will be tuned to get a suitable closed-loop response by
using auto-tune MATLAB function. Once suitable gains are
found, the transfer function block, as in Fig. 6 will be replaced
with a nonlinear mathematical model that represents an actual
vehicle, as shown in Fig. 1. In this stage, the controller should
produce almost the same result as the previous set-up.
Fig. 4. Vehicle dynamic block diagram.

2.2. System Identification

The main objective of system identification is to develop a

simplified transfer function that can be used to tune the
controller for the Adaptive Cruise Control application. A Fig. 6. PID control structure.
MATLAB system identification toolbox is utilized for this
task. A suitable model order or several poles and zeros in the
An FTP75 standard drive cycle will be used as the desired
transfer function must be carefully selected. Although there
velocity to evaluate the controller performance. The drive
are two inputs to the system, only a single input is considered.
cycle is an excellent tool to assess car performance, where the
As shown in Fig. 5, throttle input is regarded as positive input,
profile is developed to represent the actual driving behavior
while brake input is considered negative input, and the output
over an intended area.
will be the vehicle speed. Different input excitation ranges
will be used to simulate the response. The records will be used
3. Simulation Result
as training data and validation data. This profile will be
imported to the system identification toolbox to estimate a 3.1. System Identification Model Analysis
simplified transfer function based on a user preference.
As discussed in the previous chapter, although there are
two inputs to regulate the vehicle’s speed, a combined single
input between throttle and brake will be used to simplify the
identification process since a driver will never press the two
pedals at the same time. Fig. 7 shows the input u1 and output
y1 training signals, where the multi-step input signal is exited
between -100 to 100 % for the throttle and brake positions.

Fig. 5. System identification block diagram.

To measure the accuracy of the identified model, a

percentage of fits will be used to indicate whether the
performance is good or not. The cost function Mean Squared
Error (MSE) is used to determine the goodness of fit by
finding the error between the estimated output x with
reference output xref divided by the number of sample Ns as Fig. 7. Input and output signals with the excitation
given in Eq. (5). A higher percentage of fit means that the range. in between -100 to 100%.
transfer function output is very close to the actual output from
the plant.
3

Authorized licensed use limited to: Purdue University. Downloaded on February 18,2025 at 15:41:31 UTC from IEEE Xplore. Restrictions apply.
 percentage is not that significant for all the transfer functions.
Three transfer function structures with first, second and Thus, the simplest form of the first-order transfer function is
third order are considered to capture the system dynamics. It selected to keep the designing process simple, where the
can be seen from Fig. 8 that the percentage of fit for the equation is given as:
training data itself is not quite good. The first-order transfer
଴.଴଴ସଷ଻଺
function tf1 provides the highest fit (53.66%), followed by the ‫= )ݏ(ܩ‬ (6)
௦ା଴.଴ଵଵ଻ହ
second order tf2 (24.97%) and third-order tf3 (22.35%). With
new validation data, the performance of all three transfer
functions drastically drops: tf1 to -3.55%, tf2 to -87.21%, and
tf3 to -94.02%. One of the main challenges in identifying this
dynamic is that the speed will settle at 0 m/s when the brake
pedal is pushed. This behavior is difficult to capture,
especially when a large excitation is introduced. As observed,
the identified model tends to produce negative speed instead
of 0 m/s when brake is activated in both training and
validation data sets.
Fig.10. Percentage of fit for various transfer functions
with -10% to 100% excitation range.

3.2. Controller Validation with Standard Drive

Cycle

Once a suitable transfer function is obtained, the PID

controller is tuned to track the desired speed according to the
standard drive cycle. An auto-tune PID toolbox is used to get
Fig.8. Percentage of fit for various transfer functions a suitable gain. With the specification of the rise time of 8.69
with -100 to 100% excitation range. s, an overshoot of 9.34% and settling time of 55.1 s, as shown
in Fig. 11, the toolbox automatically specifies the value of
proportional gain Kp to 42.136, integral gain Ki to 2.175 and
One possible solution is constraining the output to 0 m/s derivative gain Kd to 32.133.
when designing or tuning a controller. However, even a
simple PID design rule may become complicated when
dealing with system constraints. Another alternative is to
reduce the training data’s excitation range, as shown in Fig.
9. Instead of exciting the input from the maximum value, the
second training input range u1 is excited between -10 to
100%, so the output velocity y1 did not go to 0 m/s.

Fig.11. Response of the tuned PID controller.

The PID controller is then implemented into the actual car

model system. The desired velocity is set to be the FTP75
drive cycle to evaluate the system’s effectiveness when
operating in a real driving scenario. Fig. 12 shows the control
effort of the vehicle, where it can be observed that an
unrealistic input, nearly 100% throttle position, is demanded.
The main reason behind this is due to the over-simplification
of the powertrain dynamics in the high-fidelity mathematical
Fig.9. Input and output signals with the excitation range model that relates the pedal position to torque command, as
are between -10% and 100%. discussed in Section 2.1. In reality, the relationship will be
As a result, Fig. 10 shows a higher percentage of fit for more complex as the dynamics also need to include changing
both training data with tf1 77.01%, tf2 81.86%, tf3 75.47% gear ratio, slip angle, coefficient of frictions and other
and validation data with tf1 78%, tf2 75.94%, tf3 79.2%. For parameters. Nevertheless, this work just assumed that the
this case, it can be noted that the difference between the fit plant represents a real car or maybe a low-powered car where
it requires full throttle pressing to achieve a certain desired
4

Authorized licensed use limited to: Purdue University. Downloaded on February 18,2025 at 15:41:31 UTC from IEEE Xplore. Restrictions apply.
speed. be applied.

Acknowledgement
Authors would like to acknowledge International Islamic
University Malaysia Research Management Center Grant
2020 (RMCG20-010-0010) for funding this work.

References

[1] Park, S., Rakha, H., Ahn, K., and Moran, K.: 'Fuel
economy impacts of manual, conventional cruise
control, and predictive eco-cruise control driving',
Fig.12. Demanded brake and throttle position to tack International Journal of Transportation Science and
the drive cycle. Technology, Sep. 2013, 2, (3), pp. 227–242
[2] Oguntosin, V., and Olasina, J.: 'Control design and
analysis of cruise control system', Journal of Research
Fig. 13 shows the system’s response based on the FTP75 in Engineering and Applied Sciences, Jul. 2021, 6, (3),
pp. 108–112
drive cycle. It is shown that the system (red line) manages to
[3] Lee, M. H., Park, H. G., Lee, S. H., Yoon, K. S., and
track the FTP75 drive cycle (blue line), but not as accurate as
Lee, K. S.: 'An adaptive cruise control system for
expected. The get a close behavior, more control effort is autonomous vehicles', International Journal of
needed. Since the system is physically constrained to 100% Precision Engineering and Manufacturing, Mar. 2013,
opening, the best decision that the controller can make is to 14, (3), pp. 373–380
take a longer time to reach the desired value. Again, this can [4] Zainuddin, M.A.S., Abdullah, M., Ahmad, S., and
be linked to the oversimplification of the power train Tofrowaih, K.A.: 'Performance Comparison Between
dynamics in the plant, where it needs to be more realistic for Predictive Functional Control and PID Algorithms for
future work. Automobile Cruise Control System', International
Journal of Automotive and Mechanical Engineering,
Mar. 2022, 19, (1), pp. 9460–9468
[5] Nie, Z., and Farzaneh, H.: 'Adaptive Cruise Control for
Eco-Driving Based on Model Predictive Control
Algorithm', Applied Sciences, Jul. 2020, 10, (15), pp.
5271
[6] Takahama, T., and Akasaka, D.: 'Model Predictive
Control Approach to Design Practical Adaptive Cruise
Control for Traffic Jam', International Journal of
Automotive Engineering, 2018, 9, (3), pp. 99–104
[7] Rajamani, R.: 'Vehicle dynamics and control', (New
York: Springer, Cop, 2012)
Fig.13. Response of the system based on the FTP75 [8] Vicente, B. A. H., James, S. S., and Anderson, S. R.:
drive cycle. 'Linear system identification versus physical modeling
of lateral–longitudinal vehicle dynamics', IEEE
Transactions on Control Systems Technology, May
4. Conclusion 2021, 29, (3), pp. 1380–1387
[9] Dias, J. E. A., Pereira, G. A. S., and Palhares, R. M.,:
'Longitudinal Model Identification and Velocity
This paper has presented the methodology for a system Control of an Autonomous Car', IEEE Transactions on
identification to develop longitudinal vehicle dynamics for Intelligent Transportation Systems, 2014, pp. 1–11
ACC applications. It is found that the training input range [10] Ringwood, J. V., Davidson, J., and Giorgi, S.:
plays vital role in the identification process. With a suitable 'Identifying Models Using Recorded Data', Numerical
range, the model can be identified with a good percentage of Modelling of Wave Energy Converters, 2016, pp. 123–
fit for both the training and validation data set. Although the 147
identified model can be used to tune the PID controller, [11] Keesman, K. J.: System Identification: 'An
expensive control effort may be demanded due to the limited introduction. London' (Springer, Cop, 2011)
capability of the original plant itself. Theoretically, the [12] ‘Vehicle Modeling Using Simulink’,
system identification model does its job. However, the plant https://www.mathworks.com/videos/matlab-and-
dynamics need to be improved to obtain a more realistic simulink-racing-lounge-vehicle-modeling-part-1-
result. The modelling process should include several essential simulink-
dynamics such as changes in gear ratio, differential, and other 1502466996305.html?s_tid=srchtitle_vehicle%20mo
coefficients. Future work will look at the identification delling_1, accessed 27 May 2022
process from an actual car where the same methodology can

Authorized licensed use limited to: Purdue University. Downloaded on February 18,2025 at 15:41:31 UTC from IEEE Xplore. Restrictions apply.